Consciousness, Intrinsic Motivation and Empowerment: You Can Do What You Want, but You Cannot Want What You Want

Speaker: Christoph Salge, University of Hertfordshire / New York University

Abstract:

What makes a mind want to do things in the first place? AI research focuses a lot on how to achieve a given goal or optimize for a given utility function. But how are these goals acquired in the first place, and are all possible utility functions suitable to motivate an AI towards more complex an interesting behaviour? It is hard to imagine that one could build a conscious agent that has no genuine goals or essential motivations.

The concept of intrinsic motivation (IM), an idea that originated in developmental psychology, sheds some light on these questions. Computational approaches to intrinsic motivation, which try to capture the essential driving forces for agency, have been applied to robotics, machine learning and artificial life systems.

In this talk, I will outline the properties of various computational approaches and discuss their benefits. I will then focus on empowerment, and intrinsic motivation that captures how much an agent is in control of the world it can itself perceive. I will show some examples from recent research that illustrate the behaviour that can be generated with the empowerment formalism, and outline some practical applications.











When: Wed., Jan. 31, 2018 at 6:00 pm - 7:30 pm
Where: WeWork Park South
110 E. 28th St., 8th Floor Lounge

Price: Free
Buy tickets/get more info now
See other events in these categories:

Speaker: Christoph Salge, University of Hertfordshire / New York University

Abstract:

What makes a mind want to do things in the first place? AI research focuses a lot on how to achieve a given goal or optimize for a given utility function. But how are these goals acquired in the first place, and are all possible utility functions suitable to motivate an AI towards more complex an interesting behaviour? It is hard to imagine that one could build a conscious agent that has no genuine goals or essential motivations.

The concept of intrinsic motivation (IM), an idea that originated in developmental psychology, sheds some light on these questions. Computational approaches to intrinsic motivation, which try to capture the essential driving forces for agency, have been applied to robotics, machine learning and artificial life systems.

In this talk, I will outline the properties of various computational approaches and discuss their benefits. I will then focus on empowerment, and intrinsic motivation that captures how much an agent is in control of the world it can itself perceive. I will show some examples from recent research that illustrate the behaviour that can be generated with the empowerment formalism, and outline some practical applications.

Buy tickets/get more info now